🚀 Announcing Triton 3.2.0 Windows-Native Build – Fully MSVC & NVIDIA Optimized! 🎉

[leomaxwell973]

🚀 Announcing Triton 3.2.0 Windows-Native Build – Fully MSVC & NVIDIA Optimized! 🎉

Hey everyone,

I’m excited to share a fully native Windows build of Triton 3.2.0—no WSL, no MinGW, no Linux subsystems—just pure Windows + NVIDIA compatibility. If you’ve ever wanted a truly native Triton build for CUDA development without AMD overhead or POSIX dependencies, this is it!

---

🔥 What Makes This Build Special?

✅ 100% Windows-native build – No Linux shell emulation, no virtualized dependencies.
✅ Compiled with MSVC – Uses Microsoft’s compiler instead of GCC/Clang workarounds.
✅ Stripped of AMD – AMD HIP and POSIX dependencies fully removed for pure CUDA optimization.
✅ Proton remains active – Ensuring compatibility while still being NVIDIA-exclusive.
✅ Built on Windows 11 Insider (RTX 3060, CUDA 12.1, Python 3.10.6).
✅ Optimized for portability – Debug symbols (.pdbs) and unnecessary build artifacts removed.
✅ Based on Triton’s official LLVM build – Uses the hidden Windows blob repo version of LLVM.
✅ Tweaks to NVIDIA driver.py and runtime build – Replaced POSIX-only interfaces with Windows alternatives.
✅ Full audit results – Passed all standard tests, plus 86/120 focus tests (34 expected AMD-related failures).

---

📥 Get the Build

📦 Prebuilt Wheel (PIP-Compatible)
You can grab the release directly:

pip install https://github.com/leomaxwell973/Triton-3.2.0-Windows-Nvidia-Prebuilt/releases/latest/download/Triton-3.2.0-cp310-cp310-win_amd64.whl

Or download the repository:

https://github.com/leomaxwell973/Triton-3.2.0-Windows-Nvidia-Prebuilt

---

💡 Why This Matters?

Until now, Windows users have struggled with Triton’s reliance on Linux-based tooling, and existing Windows builds required unwanted AMD support, POSIX dependencies, or hybrid environments like WSL.

• This fully native Windows build solves those issues by making Triton truly Windows-first.
• If you’re working on CUDA-accelerated AI models, this is the cleanest, fastest, and most stable Triton release you can use.

---

🔧 Known Limitations

• No AMD HIP support – This build is CUDA-exclusive, and AMD-specific tests will fail by design.
• CUPTI tweaks – Had to replace aligned_alloc with _aligned_malloc, but no confirmed performance impact yet.
• If your system somehow runs "NVIDIA HIP" (unlikely), this may not work properly.

---

💬 Final Thoughts

This has been a huge project to get Triton working in a true Windows-native way, and I’d love to hear your feedback!

🔥 Give it a spin, let me know if you run into any issues, and let’s push Triton forward for Windows users! 🚀

💡 Repo & Full Details: https://github.com/leomaxwell973/Triton-3.2.0-Windows-Nvidia-Prebuilt